Deep Learning-Based Analysis of Retinal Images for Estimation of Cardiovascular Disease Risk: the Multi-Ethnic Study of Atherosclerosis (MESA)

Background Deep learning retinal scores for cardiovascular risk have been validated only in East Asian and European-ancestry cohorts. We evaluated one in a US multi-ethnic cohort with adjudicated events and measured coronary artery calcium (CAC). Methods RetiCAC is a deep learning score estimating the probability of CAC presence from retinal photographs. In Multi-Ethnic Study of Atherosclerosis (MESA) participants without baseline atherosclerotic cardiovascular disease (ASCVD), RetiCAC was categorized by quartiles (Q1 low, Q2–Q3 moderate, Q4 high) and by fixed cut points (30.9 and 69.0) derived against measured CAC without reference to outcomes, with pre-specified hypotheses on observed 10-year risk. A retinal 10-year ASCVD risk estimator combining RetiCAC with age and sex was applied after baseline-hazard recalibration. We assessed stratification (adjusted Cox models), discrimination (c-index), reclassification (continuous NRI), calibration overall and by race/ethnicity (modified Hosmer–Lemeshow [HL] test), and residual association beyond CAC. Results Among 4,890 participants (mean age 63±10 years; 52.2% female; White, Black, Hispanic, and Chinese; follow-up 9.0±2.3 years), 9.7% had an event. Among those not high risk by Pooled Cohort Equations (PCE; n=3,469), high-risk RetiCAC had adjusted HR 4.17 (95% CI 2.24–7.76) versus low, comparable to CAC >100 (HR 3.85, 95% CI 2.70–5.50); c-indices were 0.708 (0.675–0.742) and 0.731 (0.699–0.763; p=0.098). Fixed cut points separated groups with observed 10-year risks of 2.3%, 6.6%, and 15.7% (c-index 0.712), meeting both hypotheses. Added to PCE, RetiCAC raised the c-index by 0.015 (95% CI 0.008–0.024; NRI 0.26); CAC by 0.046 (0.034–0.061; NRI 0.54), with similar gains beyond PREVENT. Residual RetiCAC not explained by CAC remained associated with events (HR 1.33 per SD, 95% CI 1.13–1.56). The retinal estimator was calibrated after recalibration (intercept −0.46, slope 0.74; HL p=0.044) and in all four race/ethnicity groups (HL p=0.14–0.96), with a c-index of 0.705 (0.684–0.727), comparable to PCE. Conclusions In a US multi-ethnic cohort, RetiCAC stratified ASCVD risk among individuals at low or moderate risk by PCE, with fixed thresholds that identified a group at high observed 10-year risk and carried prognostic information beyond CAC itself. Retinal photographs, already acquired at scale in eye care, can support cardiovascular risk stratification where CT calcium scoring is unavailable, pending prospective evaluation of these thresholds.